A Hybrid Metaheuristic based Feature Selection Framework for In-silico Mutagenicity Prediction

Abstract

Mutagenicity is both a toxic risk to humans and an indicator of carcinogenicity. Hence, estimating mutagenicity in the early stages of drug design is crucial to minimize last-stage failures and withdrawals in drug discovery. Recently, in-silico methods have started to play critical and essential roles in the drug development process because they are low cost and low effort procedures. This study aims to predict mutagenicity of chemicals using in-silico methods. To achieve this goal, a two-phased flexible framework was proposed: 1) searching the effective and representative descriptors subset with Butterfly Optimization Algorithm (BOA) and Particle Swarm Optimization and 2) predicting mutagenicity of chemicals by the selected descriptor using gradient boosted tree-based ensemble methods. The study used two datasets: one including 8167 compounds for descriptor selection and modelling, and another containing 716 external compounds to validate the efficacy of our models. The datasets comprise 162 descriptors calculated using PaDEL. The results of both the cross-validation and the external data showed that descriptors reduced by nearly one-third by BOA (51 descriptors) yielded similar or slightly better predictive results than results obtained with the entire data set. The accuracy range attained by the proposed approach using BOA is approximately 91.9% to 97.91% for the external set and 83.35% to 86.47% for the test set. This research contributes that using optimization techniques for improving early drug design and minimizing risks in drug discovery can be considered as a valuable insights and advances in the field of drug toxicity prediction, based on the findings.